As the high performance computing (HPC) moves towards AI-driven applications, these AIHPC applications demand steady and high-throughput relational data acquisition from data sites, which gradually dominates the whole computing process. In order to provide an energy-efficient service in a multi-tenant environment, we have to reduce the power consumption from serving relational data, such that the energy cost can be best amortized. While modern hardware provide multiple power modes, these power performance tradeoffs cannot be directly applied to data services due to the lack of understanding from the operation behavior under different power modes. To address this challenge, we provide an system identification study on learning the power behavior under serving database operations under different hardware modes, and propose a Control framework for Relational Operations to save Power. In contrast to today’s heuristic-based power tuning techniques, our solution achieves the goal via two facilities: (1) a control-theory based controller design that minimizes overshoot and guarantees the minimum settling time, thus control accuracy and system stability; (2) a fuzzy classifier inside database engine that helps to understand software behavior, in order to tune the sensitivity of the whole system control. We prototyped Crop as wrapping these functions in a container hierarchy, and evaluate it with workloads generated from various database benchmarks. The results show that Crop achieves up to 51.3% additional energy savings despite runtime workload dynamics and model errors, as compared to other competing methods.